Large-Scale Cross-Modal Hashing with Unified Learning and Multi-Object Regional Correlation Reasoning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep cross-modal hash retrieval (DCMHR) methods face limitations in region-specific object classification and unified hash code learning.
• Existing DCMHR models often fail to efficiently train unified hash codes or consider inter-region object correlations.

Purpose of the Study:

• To propose a novel DCMHR method, HUMOR (Large-Scale Cross-Modal Hashing with Unified Learning and Multi-Object Regional Correlation Reasoning), addressing current limitations.
• To enhance retrieval efficiency and accuracy by incorporating unified learning and multi-object regional correlation reasoning.

Main Methods:

• HUMOR utilizes Multiple Instance Learning (MIL) for reasoning label correlations within image regions.
• It employs a "reduce-add" mechanism for label rectification based on global or regional precedence.
• Unified learning of hash and classification losses is achieved through a four-step iterative algorithm for optimizing hash codes.

Main Results:

• Experiments on two baseline datasets demonstrate that HUMOR achieves higher average performance compared to most existing DCMHR methods.
• The proposed method effectively reduces model bias through unified hash code optimization.

Conclusions:

• HUMOR presents an effective and innovative approach to large-scale cross-modal hashing.
• The method significantly improves retrieval performance by addressing limitations in current DCMHR techniques.